A polishing monitoring method for polishing computation edge data

Abstract

The present application relates to the technical field of polishing monitoring, and discloses a polishing monitoring method for polishing calculation edge data.The method comprises collecting multi-dimensional polishing data streams of vibration spectrum, temperature gradient distribution and surface roughness sampling sequence, generating a polishing feature fragment set with time sequence labels through time domain and frequency domain joint segmentation.According to the polishing process standard library, the abnormal judgment threshold is matched, the abnormal segment to be verified is labeled, the local weight coefficient is calculated through the dynamic weight distribution network, the polishing quality evaluation matrix is obtained by weighting and aggregating the abnormal segment, the edge side control instruction set is generated by screening the core monitoring index, the execution priority is distributed in the cluster, and the polishing equipment group is sequentially issued.The method relies on edge computing to realize data processing nearby, improves the real-time monitoring and instruction response efficiency, accurately captures the polishing process abnormalities, and optimizes the polishing quality control process.

Description

Technical Field

[0001] This invention relates to the field of polishing monitoring technology, specifically a polishing monitoring method for calculating edge data during polishing. Background Technology

[0002] In industries such as machinery manufacturing and precision machining, polishing is a crucial process for improving the surface quality of workpieces. Its processing accuracy directly affects product performance, lifespan, and market competitiveness. As the manufacturing industry transforms towards intelligent and refined processes, traditional polishing monitoring methods are increasingly revealing their limitations. Early polishing monitoring relied heavily on operator experience, involving visual observation of the workpiece surface or tactile assessment of equipment operation. This method is highly subjective, easily influenced by human factors, and struggles to accurately identify subtle anomalies during the polishing process. Furthermore, it cannot achieve continuous, 24 / 7 monitoring, often leading to missed or incorrect assessments of quality defects, resulting in wasted raw materials and increased production costs.

[0003] With the development of sensor technology, some enterprises have introduced polishing monitoring systems based on centralized computing. These systems collect operational data by installing sensors on polishing equipment and then transmit the data to a remote cloud platform for processing and analysis. However, this centralized architecture faces challenges such as high data transmission pressure and high processing latency. The polishing process generates a massive volume of multi-dimensional data, and long-distance transmission can easily cause network congestion, leading to data loss or delays. This makes it difficult to issue monitoring commands in a timely manner and to quickly intervene in polishing anomalies. Furthermore, centralized platforms have poor adaptability to different types of polishing equipment, and anomaly detection thresholds are often fixed values ​​that cannot be dynamically adjusted according to specific process requirements, making them unsuitable for diverse polishing scenarios.

[0004] Existing monitoring technologies have significant shortcomings in data processing. Most systems analyze only single-dimensional data, lacking the ability to collaboratively process multi-dimensional data such as vibration, temperature, and surface roughness, thus failing to comprehensively reflect the complex state of the polishing process. After anomaly identification, the processing of abnormal segments lacks a scientific weighting mechanism, simply aggregating abnormal information, leading to significant deviations in quality assessment results and an inability to accurately pinpoint the core factors affecting polishing quality. Furthermore, the generated control commands lack priority classification, easily causing resource conflicts during execution, affecting the orderly and timely adjustment of equipment, further restricting the stability of polishing quality and the improvement of processing efficiency. These problems make existing polishing monitoring technologies insufficient to meet the demands of modern precision machining for high-quality and high-efficiency monitoring, necessitating a novel monitoring method to overcome these technological bottlenecks. Summary of the Invention

[0005] The purpose of this invention is to provide a polishing monitoring method for polishing calculation edge data, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a polishing monitoring method for polishing calculation edge data, the method comprising: The multi-dimensional polishing data stream of the polishing equipment group is collected, including vibration spectrum, temperature gradient distribution and surface roughness sampling sequence; The multi-dimensional polishing data stream is jointly segmented in the time domain and frequency domain to generate a set of polishing feature fragments with time-series labels; Based on the polishing process standard library, the anomaly determination threshold of the polishing feature fragment set is matched, and feature fragments exceeding the anomaly determination threshold are marked as anomalies to be verified. A dynamic weight allocation network is constructed, and local weight coefficients are calculated based on the spatial distribution density and time duration of the anomaly segment to be verified. The local weight coefficients are used to weight and aggregate the abnormal segments to be verified to generate a polishing quality evaluation matrix; Core monitoring indicators are selected based on the dimensional correlation of the polishing quality assessment matrix, and an edge-side control instruction set is generated based on the core monitoring indicators. Within the edge computing node cluster, instruction execution priorities are allocated, and the edge-side control instruction set is sent to the polishing equipment group in priority order.

[0007] Preferably, the joint time-domain and frequency-domain segmentation of the multi-dimensional polishing data stream includes: Extract the resonance peak offset sequence of the vibration spectrum and simultaneously obtain the coordinates of the spatial abrupt change points of the temperature gradient distribution; The resonance peak offset sequence is spatiotemporally aligned with the coordinates of the spatial abrupt change point to establish a vibration-temperature coupling feature mapping table. The polishing stage intervals are divided according to the discreteness of the surface roughness sampling sequence, and the effective segment of the vibration-temperature coupling feature mapping table is extracted based on the polishing stage intervals. Wavelet packet decomposition is performed on the effective segment to generate a frequency band energy proportion vector, which serves as the basic element of the polishing feature segment set.

[0008] Preferably, the anomaly determination threshold for matching the polishing feature fragment set according to the polishing process standard library includes: Load the frequency band energy percentage vector distribution range of historical qualified samples from the polishing process standard library; Calculate the Mahalanobis distance between the frequency band energy percentage vector of the current polishing feature fragment set and the historical qualified samples; When the Mahalanobis distance exceeds the preset process tolerance limit, the dynamic adjustment mechanism of the anomaly determination threshold is activated; The anomaly detection threshold is linearly compensated based on the real-time load rate and material hardness parameters of the polishing equipment group.

[0009] Preferably, constructing the dynamic weight allocation network includes: Establish a spatiotemporal distribution topology graph of the anomaly segment to be verified, where nodes represent the spatial coordinates of the anomaly segment and edges represent the temporal continuity of the anomaly segment; Calculate the degree centrality and proximity centrality of each node in the spatiotemporal distribution topology graph, and generate a node importance score; Based on the node importance score and the mechanical structure constraints of the polishing equipment group, the decay function of the local weight coefficient is derived.

[0010] Preferably, the generation of the polishing quality evaluation matrix includes: The weighted aggregation results of the abnormal segments to be verified are arranged by polishing equipment number to form the initial evaluation vector; Singular value decomposition is performed on the initial evaluation vector, and the left singular vector is extracted as the quality principal component feature; The final form of the polishing quality evaluation matrix is ​​output by performing a Hadamard product operation on the principal component features of the quality and the preset weights of the core monitoring indicators.

[0011] Preferably, the step of selecting core monitoring indicators based on the dimensional correlation of the polishing quality assessment matrix includes: Establish the Gram matrix of the polishing quality evaluation matrix and calculate its eigenvalue decay rate; When the eigenvalue decay rate is lower than a preset sensitivity threshold, the eigenvector corresponding to the largest eigenvalue is selected from the Gram matrix; The absolute values ​​of each component in the feature vector are sorted, and the polishing parameters corresponding to the top K components are selected as the core monitoring indicators.

[0012] Preferably, the generation of the edge-side control instruction set based on the core monitoring indicators includes: Based on the real-time fluctuation range of the core monitoring indicators, determine the adjustment step size and duration of the control commands; The adjustment step size and action duration are combined and optimized using fuzzy inference rules to generate a primary control command sequence; The primary control command sequence is subjected to command conflict detection, and redundant commands with overlapping spatial positions are eliminated to form the edge-side control command set.

[0013] Preferably, the allocation of instruction execution priority within the edge computing node cluster includes: Construct a directed graph of the dependency relationships of the edge-side control instruction set, where nodes represent control instructions and edges represent the sequential constraints between instructions; The directed graph of the dependencies is topologically sorted to generate the basic framework for the instruction execution sequence; Based on the real-time operating data of the polishing equipment group, a dynamic preemptive priority marker is inserted into the basic framework.

[0014] Preferably, the step of issuing the edge-side control command set to the polishing equipment group in priority order includes: The instruction sequence carrying the dynamic preemptive priority flag is split into micro-batch processing units; An instruction cache queue is established within the edge computing node cluster, and the micro-batch processing units are sorted according to their priority scores. An asynchronous dual-channel transmission mechanism is used to distribute the sorted micro-batch processing units to the corresponding polishing equipment controller.

[0015] Preferably, the method further includes: Deploy a polished knowledge graph update module in the edge computing node cluster. This module performs the following operations: Analyze the implicit process association rules in the polishing quality evaluation matrix; When a new association rule not recorded in the polishing process standard library is detected, a graph node splitting operation is triggered; The newly generated graph nodes are embedded into the topology of the original polished knowledge graph using an incremental learning algorithm.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By deploying monitoring logic on an edge computing node cluster, polishing data is collected and processed locally, effectively avoiding the latency issues caused by long-distance data transmission in traditional centralized computing architectures. The real-time data channel established between the edge computing node cluster and the polishing equipment group ensures efficient transmission of multi-dimensional polishing data streams, reduces network bandwidth consumption, avoids data loss or corruption during transmission, and allows data processing and command generation to be completed quickly at the edge, close to the equipment. This significantly improves monitoring response speed and creates conditions for timely handling of anomalies during the polishing process.

[0017] This method acquires multi-dimensional polishing data streams, including vibration spectrum, temperature gradient distribution, and surface roughness sampling sequences. Compared to single-dimensional data acquisition methods, it can more comprehensively and three-dimensionally reflect the operating status of the polishing equipment and the processing status of the workpiece, covering key parameters affecting quality during the polishing process. This provides a rich and comprehensive data foundation for subsequent anomaly identification and quality assessment. By processing the multi-dimensional data stream using joint time-domain and frequency-domain segmentation techniques, it is possible to accurately extract polishing feature segments with time-series markers, clearly presenting the dynamic changes in the polishing process, making subsequent anomaly detection more targeted.

[0018] By matching anomaly detection thresholds to a polishing process standard library, the anomaly identification criteria can be adapted to specific polishing process requirements. This overcomes the shortcomings of traditional fixed thresholds, which cannot adapt to diverse process scenarios, and improves the flexibility and accuracy of anomaly detection, enabling precise screening of anomaly segments to be verified. The construction of a dynamic weight allocation network, combining the spatial distribution density and temporal duration of the anomaly segments to be verified to calculate local weight coefficients, allows anomaly segments with different characteristics to receive differentiated attention. This avoids the evaluation bias caused by treating all anomaly segments equally in traditional methods, making subsequent weighted aggregation more scientific.

[0019] The polishing quality assessment matrix, generated by weighting and aggregating outlier segments using local weighting coefficients, systematically integrates anomaly information and clearly presents the correlations between data across various dimensions. Filtering core monitoring indicators based on matrix dimensional correlations removes redundant information, focusing on key factors affecting polishing quality. This makes the generated edge-side control command set more targeted, avoiding interference from invalid commands. Allocating command execution priorities within the edge computing node cluster and issuing control commands sequentially allows for reasonable scheduling of device resources, avoiding conflicts during command execution, ensuring orderly device adjustments, and guaranteeing the stability of the polishing process. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the working principle of the polishing monitoring method for calculating edge data in the polishing process described in this invention. Figure 2 A flowchart for the joint partitioning of the time and frequency domains; Figure 3 Flowchart for generating the polishing quality assessment matrix; Figure 4 This is a characteristic analysis diagram of singular value decomposition in polishing quality assessment. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1This invention provides a polishing monitoring method for edge data computation during polishing. The method includes: an edge computing node cluster continuously acquiring multi-dimensional polishing data streams from a polishing equipment group through a real-time data channel. These data streams specifically include vibration spectrum, temperature gradient distribution, and surface roughness sampling sequences. The acquired raw data streams are sent to a preprocessing module for joint time-domain and frequency-domain segmentation processing. The aim is to extract a set of polishing feature segments with time-series labels that characterize specific polishing stages from the continuous signal. The system accesses a pre-set polishing process standard library, which stores various characteristic parameter ranges during the processing of qualified products. The system matches the currently generated set of polishing feature segments with the benchmarks in the standard library to calculate a dynamic anomaly judgment threshold. Any feature segment exceeding this threshold is marked as an anomaly segment to be verified. To more accurately assess the impact of anomalies, the system constructs a dynamic weight allocation network. This network comprehensively considers the distribution density of the anomaly segments to be verified in the equipment space and their duration on the time axis, calculating a local weight coefficient for each anomaly segment. These weighting coefficients are used to perform weighted aggregation operations on all the abnormal segments to be verified, ultimately generating a polishing quality assessment matrix that comprehensively reflects the overall health status and processing quality of the polishing equipment group. Based on this matrix, the system further analyzes the correlation of its internal parameters and selects the core monitoring indicators that have the most significant impact on the final polishing quality. At the edge, corresponding control instruction sets are generated based on the real-time status of these core monitoring indicators. Within the edge computing node cluster, execution priorities are assigned to the instructions according to preset strategies and real-time load. The control instruction sets are then distributed to the corresponding polishing equipment groups through real-time data channels in priority order, achieving closed-loop precise control of the polishing process.

[0023] Example 1: See Figure 2The acquired multi-dimensional polishing data stream enters the joint time-domain and frequency-domain segmentation stage. This process aims to extract feature units with clear physical meaning and interrelationships from the continuous signal stream. The segmentation operation begins with in-depth analysis of vibration spectrum data, which comes from high-frequency vibration sensors installed on the polishing spindle and worktable. After smoothing and denoising the raw spectrum, the signal processing unit begins to identify and track resonance peaks in the spectrum. Resonance peaks are frequency components corresponding to local extrema on the spectral envelope, and their center frequency offsets are extremely sensitive to polishing tool wear, dynamic balance, and changes in workpiece clamping rigidity. The algorithm calculates the offset of the current resonance peak center frequency relative to the reference frequency in a fixed time window, forming a sequence of resonance peak offsets arranged in chronological order. An infrared thermal imager deployed above the polishing area continuously acquires temperature field data, generating a two-dimensional temperature gradient distribution map. The image processing algorithm performs edge detection and region segmentation on the temperature gradient distribution map, identifies pixel clusters with gradient values ​​exceeding a set threshold, and calculates the centroid coordinates of these clusters. These coordinates are defined as spatial abrupt change point coordinates, indicating the spatial location of abnormal heat points during the polishing process. Vibration and temperature signals are different physical quantities and may have different sampling frequencies, therefore, spatiotemporal alignment is required to establish a correlation. The system uses a unified high-precision timestamp to label all data and divides the polishing equipment's workspace into a grid, assigning unique coordinates to each grid. The resonance peak shift sequence and spatial abrupt change point coordinates are mapped to a unified spatiotemporal coordinate system. The frequency shift of the vibration data and the spatial coordinates of the temperature data are correlated using an interpolation algorithm to generate a vibration-temperature coupling feature mapping table. This mapping table, indexed by time, records the dominant resonance peak shift and the presence or absence of temperature abrupt changes within each spatial grid at each sampling time.

[0024] Surface roughness sampling sequences are periodically acquired by online contact or non-contact roughness measuring instruments, resulting in relatively sparse data points. To utilize surface roughness data to delineate process-significant analytical intervals, the algorithm analyzes the sequence variation trend and statistical dispersion of surface roughness sampling values. When the roughness values ​​of multiple consecutive sampling points are within a stable range and the standard deviation is below a threshold, a stable polishing stage is considered to have begun; when the roughness values ​​undergo a jump or the dispersion increases significantly, it marks the end of one stage and the beginning of another. Based on this, the entire polishing process is divided into several consecutive polishing stage intervals. Each polishing stage interval corresponds to a continuous time record in the vibration-temperature coupling feature mapping table. Data within these time intervals are extracted from the mapping table to form effective segments corresponding to specific polishing stages. Subsequent analysis will focus on these effective segments, excluding interference from irrelevant data. For each extracted effective segment, more refined frequency domain features need to be extracted; wavelet packet decomposition is applied here. Wavelet packet decomposition is a more refined time-frequency analysis method than wavelet decomposition, capable of further decomposing the high-frequency components of a signal. By selecting appropriate wavelet basis functions and decomposition levels, the vibration and temperature coupling signals (usually one-dimensional signals after fusion processing) within the effective segment are decomposed into different frequency bands. The energy of the wavelet coefficients in each frequency band, i.e., the sum of squares of the coefficients, is calculated, and the percentage of energy in each frequency band relative to the total energy is calculated, thus forming a frequency band energy proportion vector. This vector quantifies the distribution of signal energy across different frequency ranges, revealing subtle state changes during the polishing process. The set of frequency band energy proportion vectors obtained after wavelet packet decomposition of all effective segments constitutes the polishing feature fragment set; these vectors are the fundamental elements for subsequent anomaly detection.

[0025] A polishing process standard library serves as the judgment benchmark, storing a large number of historical qualified polishing process data models. When matching the anomaly judgment threshold, the system loads historical qualified sample data from the polishing process standard library that perfectly corresponds to the current polishing stage, equipment model, and material type. This historical data contains a large number of frequency band energy proportion vectors within the corresponding polishing stage. Through statistical learning, a multidimensional Gaussian distribution model of these vectors can be obtained, namely the mean vector and covariance matrix. The system calculates the Mahalanobis distance between the currently generated polishing feature segment (i.e., the frequency band energy proportion vector) and the historical qualified sample distribution. The calculation of the Mahalanobis distance requires the inverse matrix of the covariance matrix of the historical samples, which measures the degree to which the current vector deviates from the center of the historical normal distribution, while also considering the correlation between the various dimensions of the feature. The system presets a process tolerance upper limit, which is determined based on a large number of process experiments and statistical analysis, representing the allowable normal fluctuation range. When the calculated Mahalanobis distance value exceeds the preset process tolerance upper limit, the current polishing feature segment is judged to be suspicious, and the system will activate the dynamic adjustment mechanism of the anomaly judgment threshold. The dynamic adjustment mechanism incorporates real-time operating parameters of the polishing equipment group as compensation factors, primarily including the real-time load rate signal of the equipment and the material hardness parameters of the currently processed workpiece obtained through the material management system. The load rate affects the vibration and temperature rise characteristics of the equipment, while material hardness directly affects polishing force and heat generation. The adjustment algorithm linearly compensates for the initially calculated anomaly judgment threshold based on the load rate and material hardness. For example, under high load and hard material conditions, the threshold is appropriately relaxed to avoid misjudging large fluctuations under normal operating conditions as anomalies. After dynamic adjustment, a more accurate anomaly judgment threshold adapted to the current specific production conditions is finally determined. If the Mahalanobis distance of any polishing feature segment exceeds this compensated threshold, it is formally marked as an anomaly segment to be verified and enters the subsequent weight allocation and aggregation process.

[0026] Example 2: See Figure 3Constructing a dynamic weight allocation network is a crucial step in refining the processing of marked anomaly segments to be verified. The goal of the dynamic weight allocation network is to assign a local weight coefficient to each anomaly segment that reflects its relative importance. The first step in building the dynamic weight allocation network is to construct a spatiotemporal distribution topology graph of the anomaly segments to be verified. The spatiotemporal distribution topology graph is a mathematical model used to describe the spatial and temporal relationships of the anomaly segments. In the spatiotemporal distribution topology graph, each node represents a marked anomaly segment to be verified. The spatial attributes of the node are determined by the specific installation coordinates of the sensor that generated the anomaly segment on the polishing equipment, while the temporal attributes of the node are the start and end timestamps of the anomaly segment's occurrence. The connection between nodes is reflected by the presence or absence of edges. The rules for establishing edges are based on the temporal continuity and spatial proximity of the anomaly segments. If the time intervals of two anomaly segments overlap or are closely connected, and their spatial coordinates are located within the same functional unit or adjacent units of the equipment structure, then an undirected edge is established between the nodes representing these two anomaly segments. The weight of the edge can be initialized as a function value based on the time interval and spatial distance. The spatiotemporal distribution topology map constructed in this way connects discrete anomalous events into a network, revealing potential anomalous propagation paths or clustering areas.

[0027] After constructing the spatiotemporal topology graph, it is necessary to calculate the importance metric for each node in the graph. Degree centrality and proximity centrality are two commonly used indicators in network analysis. Degree centrality, for a specific node, counts the number of edges directly connected to that node. A node with a high degree centrality value indicates that it is a connection hub in the topology graph, potentially corresponding to an anomaly source with a wide range of influence. Proximity centrality is calculated by first calculating the shortest path length between any two nodes in the graph. The proximity centrality of a node is defined as the reciprocal of the sum of the shortest path lengths from that node to all other reachable nodes in the graph. A node with a high proximity centrality means that its average distance to other nodes in the network is short, and information or influence spreads from that node to the entire network faster, indicating that the node is in a central position in the network. The degree centrality and proximity centrality values ​​of each anomaly segment node in the spatiotemporal topology graph are calculated separately. These two values ​​are normalized, and a comprehensive node importance score is generated by weighted summation. The node importance score reflects the influence potential of the anomaly segment from the perspective of network topology, but the final determination of the local weight coefficients must also take into account the inherent mechanical structural constraints of the polishing equipment group. Mechanical structural constraints include physical characteristics such as the rigidity of the equipment's transmission chain, the load limits of critical bearings, and the layout of cooling pipes. For example, anomalies occurring near high-precision spindle bearings pose a far greater potential hazard than those occurring on the protective cover. The system internally stores an equipment structural criticality mapping table, which assigns different basic importance coefficients to different spatial locations of the equipment. The dynamic weight allocation network integrates node importance scores with the basic importance coefficient of the anomaly segment's location, taking into account the duration of the anomaly segment itself. Long-duration anomalies often indicate persistent rather than transient problems. Based on these factors, a decay function for local weight coefficients is derived from node importance scores, spatial location basic importance, and anomaly duration. The decay function defines how the weight changes with the spatial location of the anomaly segment (distance from the core functional unit of the equipment) and its temporal duration (length of the anomaly). It is a multivariate function expression, and the output is a unique and precise local weight coefficient for each anomaly segment to be verified.

[0028] After obtaining the local weight coefficients for each anomaly segment to be verified, the system performs a weighted aggregation operation to generate the final polishing quality assessment matrix. The weighted aggregation process categorizes and organizes the anomaly segments according to their polishing equipment numbers. The entire polishing equipment group contains multiple independently controlled polishing machines, each with a unique equipment identifier within the system. For polishing machines with the same number, all their anomaly segments to be verified are extracted. Each anomaly segment contains a quantified value representing the anomaly intensity and the previously calculated local weight coefficients. The anomaly intensity values ​​of all anomaly segments from the same machine are multiplied by their corresponding local weight coefficients, and these products are summed, or a weighted average is taken according to process requirements, to obtain an assessment value representing the overall anomaly status of that machine. This operation is repeated for all polishing machines to generate an initial assessment vector. The dimension of the initial assessment vector is equal to the total number of polishing machines, and each element in the vector corresponds to the assessment value of one machine. The initial assessment vector directly reflects the anomaly degree of each machine, but there may be strong correlations between the dimensions within the vector, and it may contain redundant information. To extract more essential and abstract quality characteristics, singular value decomposition is required on the initial assessment vector. Singular Value Decomposition (SVD) is a linear algebra tool that decomposes any matrix into the product of three matrices. Treating the initial evaluation vector as a column vector, SVD yields a left singular vector, a diagonal matrix (where the diagonal elements are singular values), and a right singular vector. The left singular vector contains the main variation patterns of the initial evaluation vector, while the magnitude of the singular values ​​in the diagonal matrix characterizes the importance of each pattern. The column vectors corresponding to the larger singular values ​​in the left singular vector are extracted; these column vectors are called quality principal component features (MPCs). MPCs capture most of the variation information of the initial evaluation vector with fewer dimensions.

[0029] The extracted principal component features are abstract features transformed mathematically, and they need to be associated with specific core monitoring indicators with clear process implications. The system pre-stores a set of preset weight vectors for the core monitoring indicators, determined by process experts based on long-term experience, reflecting the degree of influence of different monitoring indicators on the final polishing quality. The extracted principal component feature vectors are then subjected to a Hadamard product operation with the preset weight vectors of the core monitoring indicators. The Hadamard product is an element-wise multiplication operation. Through the Hadamard product operation, the mathematically extracted principal component features are "modulated" with process importance weights, making the final output polishing quality assessment matrix not only data-driven but also providing clear process guidance. The final form of the polishing quality assessment matrix is ​​a matrix with the number of rows equal to the number of principal components and the number of columns equal to the number of core monitoring indicators. This matrix provides a structured and quantitative input basis for subsequent selection of core monitoring indicators and generation of control instructions.

[0030] See Figure 4 In the polishing quality assessment process, singular value decomposition (SVD) is used for dimensionality reduction analysis and principal component feature extraction. Specifically, the polishing quality data matrix is ​​processed by SVD to generate a singular value sequence and a variance contribution rate index. The singular value distribution plot shows the distribution of singular values, with the horizontal axis representing the singular value index (1.0 to 5.0) and the vertical axis representing the singular value magnitude (0 to 3.2). The curve shows that the singular value decreases rapidly and then flattens out as the index increases, indicating that the first few singular values ​​carry the main variation information of the data. The cumulative variance contribution rate plot shows the relationship between the cumulative variance contribution rate and the number of principal components, with the horizontal axis representing the number of principal components (1.0 to 5.0) and the vertical axis representing the contribution rate. The curve shows an exponential growth trend, approaching the 0.9 threshold at the fourth principal component. In the parameter configuration, the cumulative variance contribution rate threshold is set to 0.9 to verify the rationality of retaining the first four principal components (cumulative contribution rate reaches 90%). The quality principal component features extracted by singular value decomposition provide input for the subsequent Hadamard product operation with the weights of core monitoring indicators, ultimately generating a structured polishing quality assessment matrix.

[0031] Example 3: After generating the polishing quality assessment matrix, the core task is to screen out the key monitoring indicators that have a decisive impact on polishing quality based on the internal data structure of the matrix. The screening process is strictly based on the correlation analysis between the dimensions of the matrix. The rows of the polishing quality assessment matrix typically represent different samples or time points, the columns represent different monitoring indicator parameters, and the matrix elements are quantitative assessment values ​​obtained after weighted summation and principal component extraction. Analyzing the dimensional correlations requires calculating the Gram matrix of the polishing quality assessment matrix, which is obtained by multiplying the polishing quality assessment matrix by its transpose. Gram matrix The calculation formula is:

[0032] in: This is a polishing quality assessment matrix, with the following dimensions: ( One sample, (indicators) Representation matrix transpose, It is the sample size. It refers to the number of indicators. It is The Gram matrix is ​​a square matrix. It contains information about the inner product between samples, and its eigenvalues ​​reflect the variance distribution of the original data along the directions of each principal component.

[0033] The eigenvalues ​​of the Gram matrix are calculated and arranged in descending order to form an eigenvalue spectrum. The decay rate of the eigenvalue spectrum is analyzed; the eigenvalue decay rate refers to the rate at which the ratio between consecutive eigenvalues ​​decreases. A sensitivity threshold is preset in the system; this threshold is a value between zero and one, used to determine the degree of information concentration. When the ratio of the largest eigenvalue to the second largest eigenvalue is much greater than one, and the percentage of the sum of the first k eigenvalues ​​to the total sum of all eigenvalues ​​exceeds the sensitivity threshold, the eigenvalue decay rate is considered fast, indicating that data variation is mainly dominated by the first few principal components. Conversely, if the eigenvalue decay is slow, it indicates that the information distribution is relatively dispersed.

[0034] When the eigenvalue decay rate is lower than a preset sensitivity threshold, it means that no one or a few principal components can explain the majority of the variance, requiring comprehensive screening of indicators from multiple dimensions. In this case, the algorithm selects the largest eigenvalue from the Gram matrix and extracts the corresponding eigenvector. This eigenvector is an m-dimensional column vector, where each element corresponds to the projection coefficient of a sample point in the original polishing quality assessment matrix onto the direction of maximum variance. However, our goal is to screen indicators (i.e., the columns of the matrix), not samples. Therefore, we need to shift our focus to the column space of the original polishing quality assessment matrix. The eigenvector corresponding to the largest eigenvalue of the Gram matrix, with its element weights, suggests which sample points contribute the most in the direction of maximum variance. However, to directly obtain the importance of the indicators, one practice is to calculate the eigenvector of the covariance matrix of the polishing quality assessment matrix, or to assess its importance by calculating the correlation coefficient (factor loading) between each indicator and the score of the first principal component. Specifically, the system calculates the covariance matrix of the polishing quality assessment matrix, which is an n-by-n square matrix. The eigenvalues ​​and eigenvectors of the covariance matrix are then calculated. The eigenvector corresponding to the largest eigenvalue is selected. The absolute value of each component in this n-dimensional eigenvector directly reflects the contribution of the corresponding monitoring indicator to the overall quality variation. The absolute values ​​of all n components in the eigenvector are arranged in descending order to form an ordered sequence. The top K components are selected, where K is determined by the cumulative variance contribution rate. For example, the indicators corresponding to the top K components are selected such that the variance explained by these indicators accounts for a predetermined proportion of the total variance (e.g., 85%). These K indicators are then identified as core monitoring indicators, representing the process parameters that most need to be monitored and controlled during the current polishing process.

[0035] After determining the core monitoring indicator set, the system generates an edge-side control instruction set based on the real-time data stream of these indicators. The instruction generation process begins with continuous monitoring of the real-time fluctuation amplitude of each core monitoring indicator. The fluctuation amplitude is quantified by calculating the standard deviation or range of the indicator value within a sliding time window. The magnitude of the fluctuation amplitude directly determines the adjustment step size of the control instruction. The larger the fluctuation amplitude, the more unstable the process, requiring greater intervention; therefore, the adjustment step size is set larger. Simultaneously, the algorithm analyzes the persistence of the fluctuation trend. If the indicator value continuously deviates from the target value, a longer action time is needed to correct the deviation; if it is a momentary disturbance, the action time can be set shorter. The adjustment step size and action duration are two basic parameters of the control instruction. Fuzzy inference rules are used to combine and optimize the initially determined adjustment step size and action duration. The fuzzy inference system uses the real-time fluctuation amplitude and trend direction of the core monitoring indicators as input variables. The fluctuation amplitude is divided into fuzzy sets such as "small," "medium," and "large," and the trend direction is divided into fuzzy sets such as "positive offset," "stable," and "negative offset." The output variables are the optimized adjustment step size and duration of action. The adjustment step size is divided into levels such as "fine-tuning," "medium-tuning," and "large-tuning," while the duration of action is divided into levels such as "short-term," "medium-term," and "long-term." The fuzzy inference rule base contains a series of rules in the form of "if...then...", for example: "If the fluctuation amplitude is large and the trend is positive, then the adjustment step size is large and the duration of action is medium." The system transforms the precise input values ​​into the membership degrees of fuzzy sets through a membership function, activates the corresponding inference rules, and finally transforms the inference results into precise optimized adjustment step size and duration values ​​through defuzzification. Applying these optimized parameters, a preliminary control command is generated for each core monitoring indicator, forming a primary control command sequence.

[0036] Because multiple core monitoring indicators may need to be adjusted simultaneously, and multiple instructions may be generated for the same actuator on the same polishing equipment, instruction conflict detection is necessary. The conflict detection module examines each instruction in the primary control instruction sequence, comparing the target device number, control parameter type, and actuator address space. If multiple instructions are found for the same device and actuator, and the adjustment directions are opposite (e.g., one instruction increases spindle speed, another decreases it), it is determined to be an instruction conflict. For conflicting instructions, the system arbitrates according to a predefined conflict resolution strategy based on the importance and priority of the core monitoring indicators, the expected effect strength of the instruction, and the freshness of the timestamp. Instructions with high priority, significant effect, or the latest information are retained, while redundant or contradictory instructions are eliminated. Instructions acting on different devices or different actuators are retained. After conflict detection and resolution, a coordinated and securely issued set of edge-side control instructions is ultimately formed.

[0037] Example 4: After the edge-side control instruction set is generated, execution priorities must be allocated within the edge computing node cluster to ensure that instructions are efficiently, orderly, and securely delivered to the polishing equipment. The priority allocation process begins with constructing a directed graph of the instruction set's dependencies. A directed graph of dependencies is a set of vertices and edges, where vertices represent specific control instructions, and directed edges represent the sequential execution constraints between instructions. For example, if instruction A is "stop polishing fluid supply" and instruction B is "replace polishing pad," then instruction B can only be executed after instruction A is completed. Therefore, a directed edge from vertex A to vertex B is established. The criteria for determining dependencies include the process flow diagram, equipment operation manual, and safety interlock logic. The system parses the operation object and action type of each control instruction. If instructions operate on the same physical device or involve resource contention, the order must be determined. If instructions have a logical causal relationship, such as the output of a previous instruction being a prerequisite for the execution of a subsequent instruction, a dependency edge must be established. By traversing the entire edge-side control instruction set and analyzing the potential relationships between all instruction pairs, a complete directed graph reflecting all sequential constraints between instructions is constructed. After constructing the directed graph, it needs to be topologically sorted. Topological sorting is an algorithm that arranges the vertices in a directed acyclic graph into a linear sequence, ensuring that for any directed edge from vertex u to vertex v, vertex u always appears before vertex v in the linear sequence. Topological sorting of the directed graph yields one or more valid instruction execution sequences, forming the basic framework for instruction execution order. Topological sorting algorithms are typically implemented using depth-first search or Kahn's algorithm (based on in-degree), ensuring that dependent instructions are executed in the correct process logic order. However, the basic framework generated by topological sorting is static, considering only the logical dependencies between instructions and not incorporating the dynamic real-time situation of the production site.

[0038] To cope with dynamic changes in the production environment, the system needs to incorporate real-time operating condition data from the polishing equipment group and insert dynamic preemptive priority markers into the static topology sequence. Real-time operating condition data includes equipment emergency stop signals, sensor over-limit alarms, critical component lifespan warnings, and remaining production task time. The system has a pre-defined dynamic priority rule base, which defines which types of instructions should be prioritized under what operating conditions. For example, the rules might stipulate that when an emergency stop signal is received from any equipment, all safety-related stop instructions immediately gain the highest priority and can be preemptively executed; when a sensor detects a sharp temperature rise exceeding a safety threshold, the cooling system strengthens the priority of its instructions; when the system predicts that the remaining lifespan of a critical spindle bearing is below a critical value, the related load reduction operation instructions gain higher priority. These rules are matched with the current operating condition data in real time. Once triggered, a dynamic preemptive priority marker is inserted into the basic topology sequence for the corresponding instruction, and a specific priority score is assigned. Refer to Table 1, a simplified instruction priority allocation table, which includes the instruction ID, instruction content, static topology sequence, and dynamic priority score calculated based on real-time operating conditions.

[0039] Table 1: Instruction Priority Allocation Table

[0040] During the adaptive optimization phase of instruction distribution, the complete instruction sequence carrying dynamic preemptive priority tags is broken down into smaller micro-batch processing units. The partitioning strategy for micro-batch processing units can be based on the target device of the instructions, the similarity of instruction types, or the estimated execution time window. Each micro-batch processing unit contains several logically related instructions and carries the unit's overall priority score. Within the edge computing node cluster, a global instruction cache queue is established. When micro-batch processing units to be distributed enter the queue, they are not strictly arranged according to their generation time or static topological order. The queue management mechanism dynamically sorts them according to the overall priority score of each micro-batch processing unit, with units having higher priority scores being inserted earlier in the queue, second only to units with even higher scores. This mechanism ensures that high-priority instruction groups are processed as quickly as possible. The transmission process employs an asynchronous dual-channel transmission mechanism to guarantee the reliability and real-time performance of instruction distribution. The asynchronous dual-channel transmission mechanism includes a high-priority channel and a standard-priority channel. The high-priority channel is dedicated to transmitting micro-batch processing units carrying highly dynamic preemptive priority tags. This channel may have narrower bandwidth but extremely low transmission latency and enjoys preemption rights. The standard priority channel is used to transmit all regular micro-batch processing units, with wider bandwidth, and transmits them in queue order. The two channels are physically or logically separate and operate in parallel. The instruction dispatcher determines whether to place a micro-batch processing unit into the high-priority channel queue or the standard priority channel queue based on its priority score. The high-priority channel queue is also sorted according to priority scores, and when a new high-priority unit joins, it can interrupt the standard channel transmission and immediately preempt resources to send it. This mechanism ensures both extremely low latency delivery of urgent instructions and maintains the order and stability of the regular instruction flow, ultimately delivering all instructions accurately to the corresponding polishing equipment controller for execution.

[0041] Example 5: The polishing knowledge graph update module deployed in the edge computing node cluster is responsible for enabling the system to continuously learn. The core function of this module is to analyze the data generated during system operation, automatically discover and absorb new process knowledge, thereby continuously improving the internal polishing process standard library. A polishing knowledge graph is a data model that stores and represents polishing-related knowledge in a graph structure. Nodes in the graph represent entities or concepts, such as specific process parameters, equipment components, material types, and anomaly patterns. Edges between nodes represent relationships between entities, such as "cause," "affected by," and "belong to." The polishing knowledge graph update module runs continuously in the background. Its operation is to parse the implicit process association rules in the polishing quality assessment matrix. The polishing quality assessment matrix is ​​a structured output of the system's comprehensive evaluation of the polishing process. Its row vectors represent snapshots of quality status at different time points or in different batches, and its column vectors represent different quality characteristics or monitoring indicators. Implicit process association rules refer to stable, potential association patterns between different process parameters or between process parameters and the final quality result that are not explicitly defined in advance but revealed through data analysis.

[0042] Analyzing implicit process association rules typically employs association rule mining algorithms or matrix factorization techniques from data mining. Association rule mining algorithms, such as the Apriori algorithm or FP-Growth algorithm, are used to find frequently occurring itemsets and strong association rules in the discretized data of the polishing quality evaluation matrix. For example, "when vibration feature F1 appears in the high-frequency band and temperature feature T2 exceeds the threshold in region R3, the surface roughness Ra has a high probability of exceeding the specification." Matrix factorization techniques, such as nonnegative matrix factorization, yield the polishing quality evaluation matrix as follows:

[0043] in: It is primitive Polishing quality assessment matrix It is the sample size. It refers to the number of characteristic indicators. It is A basis matrix of dimension 1. The number of latent factors is usually much smaller than , Each row of the matrix can be viewed as a representation of a sample in the latent factor space. It is A coefficient matrix of dimension 1 Each row of the matrix represents a latent factor, and each column represents the relationship between that latent factor and the original matrix. The contribution weight of each characteristic indicator. Through analysis A matrix can reveal which original feature indicators tend to be highly represented by the same latent factor, indicating a strong implicit correlation between these indicators. The constraints of nonnegative matrix factorization make the decomposition results more interpretable.

[0044] When the polishing knowledge graph update module detects a new and strong process parameter association rule through the above analysis method, and this association rule repeatedly appears in historical data to a certain level of confidence, but is not recorded in the existing polishing process standard library, the module will trigger the knowledge graph update process, specifically manifested as a node splitting operation. Node splitting is a way of knowledge graph evolution, which refers to subdividing an existing, relatively broad concept node in the graph into two or more more specific and refined sub-category nodes. For example, the original polishing knowledge graph may only contain a node named "vibration anomaly," and the rules associated with this node are relatively general. The newly discovered association rule clearly points out that when using "hard alloy" material, the vibration energy anomaly in the "2000Hz to 2500Hz" frequency band collected at the "spindle drive end bearing" location is highly correlated with the defect type "microscopic scratches on the workpiece surface." This new rule contains much more precise knowledge than the original "vibration anomaly" node. Faced with this situation, the polishing knowledge graph update module performs a node splitting operation, subdividing the original "vibration anomaly" node. The module may create new subclass nodes, such as "high-frequency vibration anomaly (driving end)" and "low-frequency vibration anomaly (non-driving end)," or create more specific nodes such as "carbide machining - high-frequency vibration anomaly at the driving end." The newly created nodes are connected to the original "vibration anomaly" parent node via edges indicating whether they are "a type" or "a subclass." Simultaneously, newly discovered strong association rules are instantiated as new edges, directly associating the newly created vibration anomaly node with the "carbide" material node, the "spindle driving end bearing" component node, and the "surface micro-scratches" defect node, assigning each edge a confidence weight. This process transforms the knowledge graph's description of process phenomena from imprecise to precise.

[0045] The newly generated graph nodes and relation edges need to be smoothly embedded into the existing polished knowledge graph topology using an incremental learning algorithm. This algorithm allows for local adjustments and optimizations of the graph using only newly discovered knowledge fragments, without requiring retraining the entire graph model. Knowledge graph embedding is a method that maps nodes and edges in a graph to a low-dimensional continuous vector space, preserving graph structure information. The incremental learning algorithm adds newly split nodes and new relation edges to the graph, forming a new, expanded graph structure. Based on the existing graph node vector representations, the algorithm learns the vector representations of new nodes by optimizing an objective function and fine-tunes the vector representations of existing nodes connected to the new edges. The objective function typically requires that the similarity between connected node vectors be higher than that between unconnected nodes.

[0046] The optimization process may employ stochastic gradient descent, updating parameters only using local graph structure data related to new and affected nodes, without loading the full graph data. Through incremental learning, the newly added "high-frequency vibration anomaly (driving end)" node acquires a vector representation in vector space that is close to the "hard alloy" material node and the "surface micro-scratches" node, while maintaining its semantic association with the parent node "vibration anomaly." Ultimately, the new knowledge is seamlessly integrated into the existing knowledge graph, updating the entire graph's topology and maintaining semantic consistency. The updated polishing knowledge graph is immediately fed back to other modules of the system, particularly the polishing process standard library and the anomaly detection module. The polishing process standard library incorporates new association rules as judgment criteria, while the anomaly detection module utilizes more refined knowledge to identify and diagnose vibration anomalies.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A polishing monitoring method for calculating edge data during polishing, characterized in that, The method is deployed in an edge computing node cluster, which establishes a real-time data channel with the polishing equipment group, and includes the following steps: The multi-dimensional polishing data stream of the polishing equipment group is collected, including vibration spectrum, temperature gradient distribution and surface roughness sampling sequence; The multi-dimensional polishing data stream is jointly segmented in the time domain and frequency domain to generate a set of polishing feature fragments with time-series labels; Based on the polishing process standard library, the anomaly determination threshold of the polishing feature fragment set is matched with the anomaly determination threshold, and feature fragments exceeding the anomaly determination threshold are marked as anomalies to be verified. A dynamic weight allocation network is constructed, and local weight coefficients are calculated based on the spatial distribution density and time duration of the anomaly segment to be verified. The local weight coefficients are used to weight and aggregate the abnormal segments to be verified to generate a polishing quality evaluation matrix; Core monitoring indicators are selected based on the dimensional correlation of the polishing quality assessment matrix, and an edge-side control instruction set is generated based on the core monitoring indicators. Within the edge computing node cluster, instruction execution priorities are allocated, and the edge-side control instruction set is sent to the polishing equipment group in priority order.

2. The polishing monitoring method for calculating edge data according to claim 1, characterized in that, The joint time-domain and frequency-domain segmentation of the multi-dimensional polishing data stream includes: Extract the resonance peak offset sequence of the vibration spectrum and simultaneously obtain the coordinates of the spatial abrupt change points of the temperature gradient distribution; The resonance peak offset sequence is spatiotemporally aligned with the coordinates of the spatial abrupt change point to establish a vibration-temperature coupling feature mapping table. The polishing stage intervals are divided according to the discreteness of the surface roughness sampling sequence, and the effective segment of the vibration-temperature coupling feature mapping table is extracted based on the polishing stage intervals. Wavelet packet decomposition is performed on the effective segment to generate a frequency band energy proportion vector, which serves as the basic element of the polishing feature segment set.

3. The polishing monitoring method for calculating edge data according to claim 2, characterized in that, The anomaly detection threshold for matching the polishing feature fragment set according to the polishing process standard library includes: Load the frequency band energy percentage vector distribution range of historical qualified samples from the polishing process standard library; Calculate the Mahalanobis distance between the frequency band energy percentage vector of the current polishing feature fragment set and the historical qualified samples; When the Mahalanobis distance exceeds the preset process tolerance limit, the dynamic adjustment mechanism of the anomaly determination threshold is activated; The anomaly detection threshold is linearly compensated based on the real-time load rate and material hardness parameters of the polishing equipment group.

4. The polishing monitoring method for calculating edge data according to claim 1, characterized in that, The construction of the dynamic weight allocation network includes: Establish a spatiotemporal distribution topology graph of the anomaly segment to be verified, where nodes represent the spatial coordinates of the anomaly segment and edges represent the temporal continuity of the anomaly segment; Calculate the degree centrality and proximity centrality of each node in the spatiotemporal distribution topology graph, and generate a node importance score; Based on the node importance score and the mechanical structure constraints of the polishing equipment group, the decay function of the local weight coefficient is derived.

5. The polishing monitoring method for calculating edge data according to claim 4, characterized in that, The generated polishing quality evaluation matrix includes: The weighted aggregation results of the abnormal segments to be verified are arranged by polishing equipment number to form the initial evaluation vector; Singular value decomposition is performed on the initial evaluation vector, and the left singular vector is extracted as the quality principal component feature; The final form of the polishing quality evaluation matrix is ​​output by performing a Hadamard product operation on the principal component features of the quality and the preset weights of the core monitoring indicators.

6. The polishing monitoring method for calculating edge data according to claim 1, characterized in that, The process of selecting core monitoring indicators based on the dimensional correlation of the polishing quality assessment matrix includes: Establish the Gram matrix of the polishing quality evaluation matrix and calculate its eigenvalue decay rate; When the eigenvalue decay rate is lower than a preset sensitivity threshold, the eigenvector corresponding to the largest eigenvalue is selected from the Gram matrix; The absolute values ​​of each component in the feature vector are sorted, and the polishing parameters corresponding to the top K components are selected as the core monitoring indicators.

7. The polishing monitoring method for calculating edge data according to claim 6, characterized in that, The generation of the edge-side control instruction set based on the core monitoring indicators includes: Based on the real-time fluctuation range of the core monitoring indicators, determine the adjustment step size and duration of the control commands; The adjustment step size and action duration are combined and optimized using fuzzy inference rules to generate a primary control command sequence; The primary control command sequence is subjected to command conflict detection, and redundant commands with overlapping spatial positions are eliminated to form the edge-side control command set.

8. The polishing monitoring method for calculating edge data according to claim 7, characterized in that, The allocation of instruction execution priority within the edge computing node cluster includes: Construct a directed graph of the dependency relationships of the edge-side control instruction set, where nodes represent control instructions and edges represent the sequential constraints between instructions; The directed graph of the dependencies is topologically sorted to generate the basic framework for the instruction execution sequence; Based on the real-time operating data of the polishing equipment group, a dynamic preemptive priority marker is inserted into the basic framework.

9. The polishing monitoring method for calculating edge data according to claim 8, characterized in that, The step of issuing the edge-side control command set to the polishing equipment group in priority order includes: The instruction sequence carrying the dynamic preemptive priority flag is split into micro-batch processing units; An instruction cache queue is established within the edge computing node cluster, and the micro-batch processing units are sorted according to their priority scores. An asynchronous dual-channel transmission mechanism is used to distribute the sorted micro-batch processing units to the corresponding polishing equipment controller.

10. The polishing monitoring method for calculating edge data according to claim 1, characterized in that, The method further includes: Deploy a polished knowledge graph update module in the edge computing node cluster. This module performs the following operations: Analyze the implicit process association rules in the polishing quality evaluation matrix; When a new association rule not recorded in the polishing process standard library is detected, a graph node splitting operation is triggered; The newly generated graph nodes are embedded into the topology of the original polished knowledge graph using an incremental learning algorithm.

Images